High-Resolution Periodical Signal Reconstruction in Power Converters Using TS-MSDA U-Net

Accurate reconstruction of high-resolution periodic signals from sparsely sampled data is critical for advanced power converters. Existing methodologies, such as coprime sampling, which redistributes samples across multiple signal periods, encounter practical challenges due to strict coprime parameter requirements. To overcome these limitations, the time-series multiscale dual-attention U-Net (TS-MSDA U-Net), an attention-based U-Net architecture, has been designed to capture complex nonlinear signal features through dual attention mechanisms. Trained with simulated data from the piecewise linear electrical circuit simulation (PLECS) platform, the model was validated on a prototype circuit and successfully reconstructed 250 kHz full waveforms with up to 36×resolution enhancement—equivalent to an effective sampling frequency of 45 MHz. Conversely, a baseline U-Net without attention mechanisms failed to reconstruct the signals, underscoring the indispensable role of attention. Both simulation and experimental results confirm the accuracy of the proposed TS-MSDA U-Net, offering a cost-effective, hardware-efficient alternative to high-speed ADCs for high-frequency signal digitization in resonant CLLC dual-active-bridge converters.